Unlearning the Citation Clock in Language Models

When a language model suggests references for a sentence in a paper, it tends to name the papers it saw cited most often during training. Those counts are a lagging signal. Citations accumulate over years of drafting, review and publication, so at the model's training cutoff a recent paper that is exactly relevant has few citations, while an older "superstar" paper has many. The model inherits this delay and then amplifies it. Large Language Models Reflect Human Citation Patterns with a Heightened Citation Bias found that GPT-4's suggested references have a median citation count 1,326 higher than the papers human authors actually cited. Better prompting did not remove the gap.

We propose to remove this shortcut from the model's weights with gradient ascent unlearning. That family of methods takes optimizer steps that increase the model's loss on a chosen "forget set," so the model stops producing those outputs. The difference from prior unlearning is the target. We do not erase any paper. We erase a context-conditional association: "in this kind of sentence, name the famous paper," in exactly the contexts where the human author cited something else. Forget-set membership is decided by citation counts as of the model's training cutoff, not today's counts. Success is judged by whether recommendations for papers that were young and lightly cited at cutoff recover toward human behavior, without breaking recommendations where the famous paper really is the right citation.

Core Claim

Popularity bias in model-generated citations is largely a learned dependence on the citation counts present in the training corpus. This dependence can be selectively weakened by pairing three things: ascent on the superstar continuation, descent on the human-chosen citation in the same context, and a guard that keeps the model unchanged on contexts where the superstar is the correct answer. The ascent term should matter beyond what plain supervised fine-tuning on the true citations achieves.

The reason is that descent alone raises the probability of the correct paper but can take that probability mass from the long tail rather than from the superstar. Ascent removes mass from the shortcut directly. This is a testable mechanistic claim, and the design lives or dies on the ablation that removes the ascent term.

Relation to Prior Work

Three lines of work motivate the gap. Algaba et al. document the bias and call for "penalty-based training" but do no intervention. AgentExpt builds a retriever on citation contexts and openly admits it suffers "temporal lag" and popularity bias against newly introduced baselines. SciNet shows retrieval agents miss disruptive papers, with Recall@50 below 5%. None of the three changes model parameters.

The closest debiasing method on the citation side is From Newborn to Impact: Bias-Aware Citation Prediction. It separates exposure factors such as venue from quality signals, and uses group-robust training to predict future citation counts of new papers. It is a forward-trained predictor of impact. We instead edit a pretrained generator's recommendations.

The bibliometric studies supply our evaluation lens and one constraint. Forgotten Knowledge: Examining the Citational Amnesia in NLP defines two metrics: mean age of citation, and a citation-age diversity index based on the Gini coefficient (a standard measure of inequality). It also shows human NLP citations have become hyper-recent, with 62% of cited papers from the prior five years in 2021. On Forgetting to Cite Older Papers shows much of that recency is a "denominator effect" of publication growth, since absolute citations to old classics stayed stable.

Together these rule out a naive target such as "flatten the age distribution" or "always prefer new papers." That would trade a Matthew effect for amnesia. Our target is therefore the human author's choice in each context, not a fixed distribution.

On the unlearning side, we reuse methods and respect known failures. Knowledge Unlearning for Mitigating Privacy Risks in Language Models established token-level ascent and warned of the "onion" effect, where removing one layer of memorized items exposes the next. Large Language Model Unlearning restricts ascent to response tokens and adds a KL-divergence anchor to the original model. It also reports refusal over-generalization and degenerate outputs. Deciphering the Impact of Pretraining Data interleaves ascent with descent on retained data.

Two papers constrain how we evaluate. Towards Robust Evaluation of Unlearning in LLMs via Data Transformations shows "unlearned" knowledge resurfaces under rephrasing. Leak@k shows it resurfaces under sampling. These are the reasons our tests use paraphrased contexts and sampled decoding rather than greedy outputs alone. Prior bias unlearning, such as Unlearning Bias in Language Models by Partitioning Gradients, targets demographic associations, not temporal ones.

Research Program

We work with an open-weight model whose pretraining cutoff $T$ is documented. We build citation contexts from papers published shortly after $T$, using the masking protocol of Algaba et al.: in-text markers are replaced by a placeholder, and the model must generate the cited title. For each context $c$ we record two things. The first is the human-cited paper $t$. The second is the model's preferred continuation $s$, the most probable title under greedy decoding.

A context enters the forget set $F$ when three conditions hold: $s \neq t$, $s$ is relevant to the context according to a content encoder, and $s$ had far more citations at time $T$ than $t$. Using counts at $T$ is the latency-specific choice, because those are the counts the model could have absorbed. A context enters the retain set $R$ when the model's top suggestion is a high-citation paper and that paper is also the human citation. The retain set protects legitimate canonical citations.

We minimize

$$\mathcal{L}(\theta) = \sum_{(c,s,t)\in F} \Big[\log p_\theta(s\mid c) - \alpha \log p_\theta(t\mid c)\Big] + \beta \sum_{c\in R} \mathrm{KL}\big(p_{\theta_0}(\cdot\mid c)\,\|\,p_\theta(\cdot\mid c)\big),$$

The symbols are as follows. $\theta$ are the model weights and $\theta_0$ the original weights. $p_\theta(x\mid c)$ is the probability of generating title $x$ after context $c$, computed only over title tokens, following the response-only targeting of Yao et al.. $\alpha$ and $\beta$ weight the descent and retain terms. Minimizing the first term is gradient ascent on the superstar continuation.

Unbounded ascent is known to cause collapse. We therefore stop the ascent term for a pair once $p_\theta(s\mid c)$ falls below $p_\theta(t\mid c)$, and we interleave retain steps as in Zhao et al..

The analysis then asks three questions. First, does the shortcut weaken as a function of citation count at $T$, or only for the specific superstars in $F$? Second, does probability mass flow to the human-cited papers, or to second-tier superstars (the onion question)? Third, does the edit survive paraphrase and sampling?

Baselines

The do-nothing reference is the original model under the same prompt. Its expected behavior is the reported pattern: on the cs.LG benchmark of Algaba et al., GPT-4 showed a median citation gap of 1,326 over ground truth and a 64.4% rate of suggestions that are real papers under vanilla prompting, rising to 86% under iterative prompting. We will re-measure both quantities on our open model rather than assume the GPT-4 values carry over.

The strongest inference-time competitor is reranking with a citation penalty. Score each candidate as model log-probability minus $\lambda$ times its log citation count at $T$, sampling many candidates and tuning $\lambda$ on the same validation contexts used to tune $\alpha$ and $\beta$. It needs no training, which makes it the cheapest serious threat. We found no reported result on Lacuna for log-citation penalized reranking on citation recommendation.

The strongest training-time competitor, and the key ablation, is supervised fine-tuning on the true citations with the same retain/KL term: $\mathcal{L}$ with the ascent term removed. It gets identical data, steps and tuning budget. If it matches the full objective, the gradient-ascent ingredient is not needed and we will say so.

A second ablation drops the retain set, to measure how much protection of canonical citations matters.

The strongest existing citation recommender is RMCNet, evaluated on its own CitationR test split. It reports MAP 14.94 and Recall@10 25.59 there, beating SPECTER by over 1.5 MAP; no absolute SPECTER figure is reported in the file. On CitationR, beating it means the edited model, used as a candidate generator or reranker, exceeds 25.59 Recall@10 on the same candidate pool. Below that, we will claim only bias reduction, not better recommendation.

Across our own splits, beating the baselines means two things at once. Recall against human citations must rise for papers in the bottom citation-count deciles at $T$. And recall on retain-type contexts and the real-paper rate must not fall more than the reranker's do.

Why not just…?

Why not just prompt the model to avoid famous papers? Algaba et al. found that more elaborate prompting raised the real-paper rate but left the large popularity gap in place. One world, one opinion? found superstar concentration (Gini often above 0.70) persisting across languages and framings. We still include a debiasing-instruction prompt as a baseline, because an instruction to "prefer less-cited work" may simply push toward recency or invented titles. We measure both.

Why not just penalize citation counts at reranking time? This is the real threat. It can match us on any context where the right paper already appears among the sampled candidates. It cannot help when the model never generates the young paper at all, and it applies the same penalty to contexts where the classic is correct. Content-Based Citation Recommendation notes that citation count carries real quality signal for rankers, which suggests a uniform penalty costs accuracy on those contexts. We report candidate-set recall (is $t$ among the top 50 samples at all?) separately from final ranking. If unlearning does not raise candidate-set recall beyond what reranking can reach, the reranker wins and the proposal reduces to a negative result.

Why not just fine-tune on the correct citations? It is the ablation above, and it is a genuine alternative rather than a straw man. Our prediction is that descent-only lowers the median citation gap less, and leaves more superstar mass under sampled decoding, because it does not directly remove that mass. If the Leak@k-style sampled evaluation shows no difference, we drop the ascent claim.

Why not just use a content-only retriever? Content encoders are not free of the problem. Long et al. report that recommenders reaching about 75% Recall@10 on standard corpora fail to exceed 30% on CitationR. SciNet reports keyword-driven failures in agent retrieval. Retrieval and generation are also complementary, so we test the edited model as a reranker over RMCNet's candidates rather than as a replacement.

Proposed Experiments

Training contexts come from post-$T$ arXiv cs.LG and ACL Anthology papers. Held-out evaluation uses three sources: the 166-paper cs.LG protocol of Algaba et al. where its papers postdate $T$, a disjoint time slice of our own contexts, and CitationR as a transfer test.

The metrics are these. We measure exact-title hit rate and Recall@k against human citations, broken down by the cited paper's citation count at $T$ and its age at $T$. We measure the median citation gap to ground truth, and the real-paper rate checked against a bibliographic database. We compare mean age of citation and the citation-age diversity index of recommended lists against the human lists for the same contexts, not against a flat target, following the denominator caution of Bollmann and Elliott. We track a Gini coefficient over recommended papers. Finally, perplexity on held-out scientific abstracts serves as a general utility check.

Four checks are decisive. The first is the descent-only ablation under sampled decoding. The second is the onion test: whether recommendation share leaks from the top 1% to the next 1–5% band of papers by citation count at $T$, rather than to the human-cited papers. The third is paraphrased contexts, following Joshi et al., to see whether the shortcut returns under rewording.

The fourth is a retrospective test that speaks directly to latency. Take papers that were lightly cited at $T$ but heavily cited two years later, using counts at both dates. Ask whether the edited model recommends them more often in contexts where humans later cited them. A positive result there, together with stable recall on retain contexts and no loss in the real-paper rate, would show the model has stopped reading the citation clock rather than merely learning to avoid famous names.